CMFDFormer: Transformer-based Copy-Move Forgery Detection with Continual Learning
The CMFDFormer network uses a Transformer-style backbone and a Pluggable Hybrid Decoder to detect copy-move forgery. A novel PCSD continual learning framework is introduced to prevent performance loss when the model encounters new forgery tasks. The system leverages intermediate features and dual pooling strategies for knowledge distillation, showing better robustness and comparable performance across different backbones.